What is the difference between a human and a program?
Neural networks, which currently constitute almost the entire field of artificial intelligence, can consider many more factors in decision-making than a human, doing so faster and, in most cases, more accurately. However, programs operate only as they have been programmed or trained. They can be very complex, taking multiple factors into account and acting in highly variable ways. Still, they cannot replace a human in decision-making. So, what distinguishes a person from such a program? Here, three key distinctions should be noted, from which all others follow:
- A human possesses a worldview that allows them to supplement their understanding with information that is not explicitly programmed. Additionally, this worldview is structured in such a way that it enables us to have at least some idea about everything, even if it is something round and glowing in the sky (UFO). Typically, ontologies are constructed for this purpose, but ontologies lack such completeness, poorly consider the polysemy of terms, their interrelations, and are still applicable only in strictly limited topics.
- A human has logic that takes this worldview into account, which we refer to as common sense. Any statement carries meaning and encompasses hidden, undeclared knowledge. Despite the fact that logical laws have existed for many centuries, no one still knows how ordinary, non-mathematical reasoning functions. Essentially, we do not know how to program even ordinary syllogisms.
- Arbitrariness. Programs do not possess arbitrariness. This is perhaps the most complex of all three distinctions. What do we mean by arbitrariness? The ability to develop new behavior different from what we performed under the same circumstances before, or to create behavior in new, previously unencountered circumstances. In other words, it is essentially the creation of a new behavior program on the fly without trial and error, taking into account new internal circumstances as well.
Randomness remains an unexplored field for researchers. Genetic algorithms that can generate new behavioral programs for intelligent agents are not the solution, as they produce results not logically but through 'mutations,' and the solution is found 'randomly' during the selection of these mutations, that is, through trial and error. A human finds the solution immediately, logically constructing it. A person can even explain why a particular solution was chosen. A genetic algorithm lacks reasoning.
It is known that the higher an animal is on the evolutionary ladder, the more random its behavior can be. The greatest randomness is exhibited in humans, as humans can consider not only external circumstances and their learned skills but also hidden factors—personal motives, previously communicated information, and results from actions in similar circumstances. This significantly increases the variability of human behavior, and in my opinion, consciousness is precisely responsible for this. But more on that later.
Consciousness and randomness
What does consciousness have to do with it? In behavioral psychology, it is known that we perform habitual actions automatically, mechanically, without the involvement of consciousness. This is a remarkable fact, indicating that consciousness is involved in creating new behavior and is linked to orienting behavior. This also means that consciousness engages precisely when it is necessary to alter a habitual behavior pattern, for example, to respond to new demands in light of new opportunities. Some scientists, such as Dawkins or Metzinger, have also suggested that consciousness is somehow connected to individuals having a self-image, meaning that the model of the world includes a model of the subject itself. So what should a system that possesses such randomness look like? What structure should it have to build new behavior in order to solve tasks according to new circumstances?
To begin with, we need to recall and clarify some known facts. All animals with a nervous system contain a model of their environment, integrated with a repertoire of actions they can take within it. This means it is not just a model of the environment, as some researchers suggest, but also a model of possible behavior in various situations. At the same time, it serves as a prediction model for changes in the environment in response to any animal actions. Cognitive scientists often overlook this, despite clear indications from mirror neurons in the premotor cortex, as well as studies observing the activation of neurons in macaques, which respond to the perception of bananas. Not only do the areas related to the banana in the visual and temporal cortices activate, but also the hand in the somatosensory cortex, because the concept of a banana is directly linked to the hand. A monkey is only interested in the fruit it can grasp and eat. We tend to forget that the nervous system did not evolve to reflect the world as perceived by animals. They are not sophists; they simply want to eat. Therefore, their model is more about behavior than about environmental reflection.
This model already possesses a certain degree of arbitrariness, reflected in the variability of behavior under similar circumstances. Animals have a repertoire of possible actions they can take depending on the situation. These can be more complex temporal patterns (conditioned reflexes) than immediate reactions to events. However, it is still not completely arbitrary behavior, which allows us to train animals but not humans.
Here, an important consideration comes into play: the more familiar the circumstances, the less variability in behavior, as the brain has a solution. Conversely, the newer the circumstances, the greater the number of possible behaviors. The whole question lies in their selection and combination. Animals achieve this by displaying their entire repertoire of possible actions, as demonstrated in the experiments of Skinner.
It cannot be said that arbitrary behavior is entirely new; it consists of previously learned behavior patterns. It is their recombination, initiated by new circumstances that do not completely match those for which there is an existing pattern. And this is precisely where the distinction between arbitrary and automatic behavior lies.
Modeling Arbitrary Behavior
Creating a program for arbitrary behavior that can account for new circumstances would make it possible to establish a universal 'program of everything' (analogous to a 'theory of everything'), at least for certain domain tasks.
What could make their behavior more arbitrary and free? My experiments showed that the only solution is the existence of a second model that simulates the first and can change it, meaning it does not interact with the environment like the first one, but with the first model to modify it.
The first model reacts to environmental circumstances. If a new pattern activated by it is found, the second model is triggered, which is trained to seek solutions within the first model by recognizing all possible behavior options in the new context. Let me remind you that in a new context, more behavior options are activated, so the question is indeed about their selection or combination. This happens because, unlike in a familiar situation, in response to new circumstances, not one behavior pattern is activated but several at once.
Every time the brain encounters something new, it performs not one, but two acts – recognizing the situation in the first model and recognizing already executed or possible actions in the second model. And it is in this structure that many possibilities emerge, similar to consciousness.
- This two-act structure allows for consideration of not only external but also internal factors – in the second model, the outcomes of previous actions, distant motives of the subject, etc., can be remembered and recognized.
- Such a system can construct new behavior immediately, without long training initiated by the environment according to evolutionary theory. For example, the second model has the capability to transfer solutions from one sub-model of the first model to other parts of it and many other possibilities of the metamodel.
- A distinguishing feature of consciousness is the awareness of its actions, or autobiographical memory, as shown in the article (1). The proposed two-act structure possesses this ability — the second model can store data about the actions of the first (no model can store data about its own actions, as it must contain coherent models of its actions rather than reactions to the environment).
But how does the construction of new behavior occur in the two-act structure of consciousness? We lack a brain or even a plausible model of it. We began experimenting with verb frames as prototypes of the models contained in our brain. A frame consists of a set of verb actors to describe a situation, and a combination of frames can be used to describe complex behavior. Situation description frames belong to the first model, while the frame that describes its actions belongs to the second model with personal action verbs. They often get mixed up, as even a single sentence is a mixture of several acts of recognition and action (speech act). The construction of long speech expressions itself is the best example of arbitrary behavior.
When the first model of the system recognizes a new pattern for which it has no programmed response, it triggers the second model. The second model gathers activated frames from the first and looks for a shorter path in the graph of interconnected frames that most effectively ‘closes’ the patterns of the new situation with a combination of frames. This is a rather complex operation and we have yet to achieve results that could be called a ‘universal program’, but the initial successes are promising.
Experimental studies of consciousness through modeling and comparing software solutions with psychological data provide interesting material for further research and allow testing some poorly validated hypotheses from experiments on humans. This can be termed modeling experiments. And this is just the initial result in this research direction.
Bibliography
1. .
Source: habr.com
